IRJET - Steering Wheel Angle Prediction for Self-Driving Cars

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 4627
STEERING WHEEL ANGLE PREDICTION FOR SELF-DRIVING CARS
Kiran U Kamath1, Dhanush B Raj1, Meghana G S1, Prajwal T1, Suhas Bharadwaj R2
1Dept. of ISE, The National Institute of Engineering, Mysore
2Asst. Professor, Dept. Of ISE, The National Institute of Engineering, Mysore
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - In this paper, a self-driving car prototype using
Deep Neural Network is proposed. Self-driving cars are
one of the most increasing interests in recent years.
Convolutional Neural Networks(CNNs) have been shown
to achieve significant performance in various perception
and control tasks compared to other techniques in the
latest years. The automated system learns to drive on the
local roads with no obstructions and also on the highways
with the lane markings on it. This is because of the
convolutional neural network (CNN) which is trained to
map the pixels from the minimum training data provided
by humans. It operates perfectly fine in the areas having
hazy visuals like parking lots etc.
With human steering angle as the limited training signal,
the system improvises and automatically learns to detect
the road visibilities. Despite no explicit training for the
detection of outline of the road, the system efficiently
optimizes all processing steps simultaneously when
compared to explicit decomposition of the problem.
It can be ensured that this will lead to better performance
with smaller systems. The overall system performance
maximizes by the optimization of system performance. The
system learns to solve the problem with a minimum number
of processing steps which results in smaller networks.
Key Words: Self-driving car, Deep Learning,
Convolutional Neural Networks, Behaviour Cloning,
Convolutional kernel, Feature map
1. INTRODUCTION
CNNs have revolutionized the computational pattern
recognition process. Before widespread adoption of CNNs,
most pattern recognition tasks were performed using an
initial stage of manual feature extraction followed by a
classifier.
A Convolutional Neural Network is a class of artificial
neural networks that uses convolutional layers to filter
inputs for useful information. The convolution operation
involves combining input data (feature map) with a
convolution kernel (filter) to form a transformed feature
map. The filters in the convolutional layers (conv layers)
are modified based on learned parameters(weights plus
bias) to extract the most useful information for a specific
task. Convolutional networks adjust automatically to find
the best feature based on the task. The CNN would filter
information about the shape of an object when confronted
with a general object recognition task but would extract
the edges of the road when faced with an edge recognition
task[1]. This is based on the CNN’s understanding that
different classes of objects have different shapes but that
different types of roads are more likely to differ in shape
than in color.
Convolutional networks are composed of an input layer,
an output layer, and one or more hidden layers. This
allows CNN to transform an input volume in three
dimensions to an output volume.
1.1 DATA COLLECTION
We have used data collected from a simulator. While
driving a car in a simulator is certainly not the same as
driving a car in the real world, there are many similarities.
Given the current state of game graphics, images captured
in a simulated environment are a good approximation of
images that could be captured in the real world. Of course,
the simulator also affords safety and ease-of-use. Data
collection is much simpler and most importantly, the failed
model could cause no threat to life. A successful model
might be implemented later in a real car with real
cameras.
Our project is mainly based on data that is collected based
on behavior cloning[3]. The goal of behavioral cloning is to
collect data while exhibiting good behavior and then train
a model to mimic that behavior with the collected data.
The images that were captured while driving will be fed
into a neural network to teach the model how to drive
properly.
Each image is captured with the accompanying steering
angle. The images will be fed into the network, and the
network’s job will be to match the appropriate steering
angle.
2. MODEL ARCHITECTURE
We train our model to minimize the mean squared error
between the steering command output by the model and
the values in the dataset[1]. The network consists of 9
layers, including a normalization layer, 5 convolutional
layers and 3 fully connected layers.
The first layer of the network performs image
normalization. Normalization normalizes the data
between 0–1 by dividing training data and testing data by
factor(mostly 255 incase of images). Training and
validation set should be normalized with the same factor.

Next 5 layers are convolutional layers which help in
feature extraction. We use strided convolutions in the first
three convolutional layers with a 2×2 stride and a 5×5
kernel and a non-strided convolution with a 3×3 kernel
size in the last two convolutional layers. Hyperparameters
like depth, stride control the size of the output volume.
When the stride is 1 then we move the filters one pixel at a
time[3]. When the stride is 2 then the filters jump 2 pixels
at a time as we slide them around. This will produce
smaller output volumes spatially. We follow the five
convolutional layers with three fully connected layers
leading to an output value of steering wheel.
Fig-1 : CNN Architecture
3. TRAINING THE MODEL
3.1 Data Selection
The first step to training a neural network is selecting
the frames to use. Our collected data is labeled with road
type, weather condition, and the driver’s activity. To train
a CNN to do lane following we only select data where the
driver was staying in a lane and discard the rest. We then
sample that video at 10 FPS.
3.2 Augmentation
After selecting the final set of frames we augment the data
by adding artificial shifts and rotations to teach the
network how to recover from a poor position or
orientation. The magnitude of these perturbations is
chosen randomly from a normal distribution[5]. The
distribution has zero mean, and the standard deviation is
twice the standard deviation that we measured with
human drivers.
We apply random shear operation to the images
generated by the simulator in the preprocessing step. This
is done basically to help the car to navigate in the training
track.
Fig-2: Sheared image after random shaer operation
The images captured by the simulator come with a lot of
unnecessary information like the sky, which is not useful
for the training of our model. So we crop the useless data
from the sheared images.
Fig-3 : Cropped image after cropping the sheared image
Next we apply a random flip operation. The idea behind
this operation is left turning bends are more prevalent
than right bends in the training track. Hence, in order to
increase the generalization of our mode, we flip images
and respective steering angles.
Fig-4 : Flipped image after flipping the cropped image.
4. SIMULATION
Before road-testing a trained CNN, we first evaluate the
performance of the network in simulation.
The simulator takes the pre-recorded from the virtual
front facing cameras of the car, on the manual driving
mode and generates the images. The simulator accesses
the recorded test video along with the accurate steering
commands during the capture of the video. The simulator
sends the first frame of the test video for any shift and
rotation and then sends to the trained CNN[4]. Then the
CNN returns the steering angle for that frame. Returned
steering angles are fed into the dynamic vehicle so that the
orientation and position of the simulated vehicle gets
updated. The simulator then modifies the next frame in
the video so that it appears as if the vehicle were at the
position that the steering angle returned by the CNN.

5. EVALUATION
We have the networks provide steering commands in
our simulator.
We estimate what percent of the time the network
could drive the car(autonamy). The metric is determined
by counting simulated human interventions. These
interventions occur when the simulated vehicle departed
from the centreline by more than 1m[3]. We assume that
in real life an actual intervention would require a total of
6second: This is the time required for a human to retake
control of the vehicle, recenter it, and then restart the self
steering mode. We calculate the percentage autonomy by
counting the number of interventions multiplying by 6
seconds, dividing by the elapsed time of the simulated test,
and then subtracting the result from 1:
(
( )
[ ]
)
Suppose, if we had 20 interventions in 1200 seconds,
we would have an autonomy value of
( )
6. CONCLUSION
We have successfully shown that CNNs are able to
understand the entire learning process lane and road
following without manual decomposition into road or lane
marking detection, path planning, and control. A small
amount of training data from less hours of driving was
sufficient to train the virtual car to operate in diverse
conditions, on highways, local and residential roads in
sunny, cloudy, and rainy conditions[4]. The CNN is able to
extract the meaningful and useful road features from a
very sparse training signal(only steering). For example
our autonomous system learns to detect the outline of a
road without the need of explicit labels during training.
As can be seen, there are many areas we could explore
to push this project further and obtain even more
convincing results. Furthermore, we are going to take
throttle into the model with the ambition of achieving
higher levels of autonomous cars.
REFERENCES
[1] Kornack and P. Rakic, “Cell Proliferation without
Neurogenesis in Adult Primate Neocortex,” Science,
vol. 294, Dec. 2001, pp. 2127-2130,
doi:10.1126/science.1065467.
[2] S. Ren, K. He, R. Girishick, and J. Sun, “Faster R-
CNN:towards real time object detection with region
proposal networks,” Advances in Neural Information
Processing Systems(NIPS), 2015.
[3] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B.
Fleep, P. Goyal, L. D. Jackel, M.Monfort, U. Muller, and J.
Zhang, “End to end learning for self-driving cars,”
2016.
[4] Advanced Applied Deep Learning ,Convolutional
Neural Networks and Object Detection, Umberto
Michelucci, Dec 26,2019
[5] Hands-On Machine Learning with Scikit-Learn and
TensorFlow by Aurélien Géron

IRJET - Steering Wheel Angle Prediction for Self-Driving Cars

Recommended

Recommended

More Related Content

What's hot

What's hot (18)

Similar to IRJET - Steering Wheel Angle Prediction for Self-Driving Cars

Similar to IRJET - Steering Wheel Angle Prediction for Self-Driving Cars (20)

More from IRJET Journal

More from IRJET Journal (20)

Recently uploaded

Recently uploaded (20)

IRJET - Steering Wheel Angle Prediction for Self-Driving Cars